First exploration of RPG dataset

Author

Romain Frelat

Published

October 15, 2025

Identification of the case study

On October 2nd, the dataset consisted of 1472 field coordinates, spread over 27 French departments. Two departments seems to offer ideal case study (Figure 1) :

  • Ille-et-Vilaine (35) with 4 different datasets and 227 points;
  • Gironde (33) with 3 different datasets, and 338 points.
Figure 1: Distribution of the fields per department represented as the number of fields (x-axis) and the number of projects per department (y-axis). The numbers shown in the plot are the number of the department.

Spatial coverage

RPG sequence de culture

Let’s try using the RPG sequence de culture from

Girault, Baptiste; Martin, Philippe, 2025, “Séquences de culture, France, 2015-2023”, https://doi.org/10.57745/VMYCYM, Recherche Data Gouv, V2

The main advantages are that (1) it is very easy to download (one file per department), (2) relatively light (100Mb per file), (3) can retrieve the full crop sequence (2015-2023) in a single operation.
The main drawbacks are that (1) it doesn’t contain RPG complété information (coordinates might fall outside the dataset) and (2) the fields are subdivided so it’s harder to get information on field size.

To get more information about the crop classes, the following companion dataset is needed:

Girault, Baptiste; Martin, Philippe, 2023, “Référentiel des cultures RPG”, https://doi.org/10.57745/FMP8GU, Recherche Data Gouv, V1

Table 1: Number of fields in Ille-et-vilaine with information from RPG Sequences de culture (TRUE) or not (FALSE)
BIOMHE BISCO DIVAG SEBIOPAG_ZAAr
FALSE 1 1 0 0
TRUE 39 26 40 120
Table 2: Number of fields in Gironde with information from RPG Sequences de culture (TRUE) or not (FALSE)
BACCHUS LepiBats OSCAR
FALSE 14 18 3
TRUE 258 28 17

The coverage of RPG sequence de culture is very good in Ille-et-vilaine (Table 1, only 2 missing fields), but less complete in Gironde (Table 2, with 35 missing fields).

RPG Complete

A team from INRAE are completing the RPG with fields that are not included originally. For each of these fields, the crop sequence is also provided since 2016.

The RPG complete 2023 is not fully released yet (missing regions), so we will explore the dataset of 2022.

Cantelaube, Pierre; Lardot, Benjamin, 2024, “RPG complété 2022 Région Nouvelle-Aquitaine”, https://doi.org/10.57745/6FNRWO, Recherche Data Gouv, V5 (d33: 47Mb compressed, 171Mb extracted)
Cantelaube, Pierre; Lardot, Benjamin, 2024, “RPG complété 2022 Région Bretagne”, https://doi.org/10.57745/DL3O6C, Recherche Data Gouv, V1 (d35, 53Mb compressed, 180Mb extracted)

Table 3: Number of fields in Ille-et-vilaine with information from RPG
BIOMHE BISCO DIVAG SEBIOPAG_ZAAr
None 1 0 0 0
RPG 39 26 40 120
RPG Complete 0 1 0 0
Table 4: Number of fields in Gironde with information from RPG Complété
BACCHUS LepiBats OSCAR
None 0 17 3
RPG 258 28 17
RPG Complete 14 1 0

The RPG complete doesn`t solve all the issues of sites not fitting within agricultural fields (Table 3, Table 4)…

A first visual exploration seems to indicate samplings from vineyard or forested areas (Figure 2, Figure 3).

Visual exploration

Figure 2: Interactive map of the fields sampled in Gironde
Figure 3: Zoom in an area with multiple fields around Langon

TO BE DISCUSSED:

  • fields in RPG sequence de culture and RPG complete seems very patchy from Figure 3, yet the coverage is not bad (>80%). Should we complete the data with other data sources, e.g. OSO land cover? If yes, how?. So far, OSO only available for 2018-2021 here, but there might exist other access. Or should we discard points with missing information / possibly incorrect coordinates?
  • this is an important question for the buffer areas (500 to 1500m). How do we handle percentage of land cover when so many areas are missing?

Crop rotation (2015-2023)

Let’s zoom in Ille-et-vilaine (35) with the RPG sequence de culture on crop rotation.

The most common crops are shown in Table 5.

Table 5: Top ten crop in the fields sampled in Ille-et-vilaine
Var1 Freq Nom
4 BTH 522 Blé tendre d’hiver
37 PTR 345 Autre prairie temporaire de 5 ans ou moins
23 MIE 299 Maïs ensilage
24 MIS 189 Maïs
30 ORH 98 Orge d’hiver
11 CZH 97 Colza d’hiver
19 MCR 78 Mélange de céréales
27 MLG 52 Mélange de légumineuses prépondérantes au semis et de graminées fourragères de 5 ans ou moins
29 MPC 38 Mélange de protéagineux (pois et/ou lupin et/ou féverole) prépondérants semés avant le 31/05 et de céréales
18 LUZ 29 Autre luzerne

The most common crop sequences are shown in Table 6.

Table 6: Top ten crop sequences in the fields sampled in Ille-et-vilaine
rotation Freq
MIS,BTH,MIE,BTH,MIE,BTH,MIE,BTH,MIE 21
PTR,PTR,PTR,PTR,PTR,BTH,MIE,BTH,MIE 8
BTH,MIS,ORH,BTH,MIS,ORH,BTH,MIS,ORH 7
CZH,BTP,MLC,CZH,MCR,MLT,CZH,SGH,CZH 7
LUZ,LUZ,LUZ,BTH,MIE,BTH,MIE,BTH,MIE 7
MIS,BTH,MIS,BTH,MIS,BTH,MIS,BTH,MIS 7
BTH,BVF,BTH,MIE,BTH,MIE,BTH,MIE,ORH 6
BTH,MIE,BTH,MIE,BTH,CZH,BTH,MIE,MIE 6
BTH,MIE,BTH,MIE,BTH,MIE,MIE,BTH,MIE 6
BTH,PTR,PTR,PTR,PTR,PTR,PTR,MCR,MIE 6

The most common crop varieties cultivated on the same fields are shown in Table 7.

Table 7: Top ten crop diversity on field (unique crop, independent of the order)
u_rotation Freq
BTH,MIE,MIS 23
BTH,MIS,ORH 12
BTH,MIE,PTR 9
BTH,MIS 9
BTH,CZH,MIE 7
BTH,CZH,MIS 7
BTH,LUZ,MIE 7
BTH,MCR,MIE,PTR 7
BTP,CZH,MCR,MLC,MLT,SGH 7
AVP,CPL,MPC,ORH,ORP,SRS,TRE,TTH 6
Figure 4: Distribution of the number of different crops cultivated per field in the period 2015-2023

On most fields, there are between 3 and 4 different crops cultivated within the 2015-2023 period.

TO BE DISCUSSED:
- decide whether we simplify the crop types (e.g. merging MIE and MIS)
- is the order of the successive crops important? (e.g. CZH,BTH,MIE different from BTH,CZH,MIE ?)

TO BE IMPROVED:
- subset the crop sequence for each field between N-5 and N.

Field size

Area

Fields in the RPG sequence de culture can be sub-divided depending in the time period 2015-2023. So we have to re-calculate the size of the entire field of the year when the study was carried out.

For the area, it is simplified thanks to the variable seq_surf giving the area of the sub-fields in ha, and parcelYYYY with the ID of the parcel for each year.

Figure 5: Correlations among field areas in the time series. The subarea is the area from sub-fields provided directly in the RPG sequence de culture. The areas are calculated from all the fields with the same ID in a given year.

TO BE DISCUSSED:
- the correlation in field areas in the successive years are (relatively) low compared to my expectations (field size shouldn’t change so much in 10-year period, and more closely related years should be more correlated). Are the results shown in Figure 5 normal or does it indicate possible errors in merging subfields?

Perimeter

This is a tricky operation, because sub-fields do not overlap properly (different edges), and sometimes they have complex geometry. So far, the best solution that I found is to:

  1. aggregate (or dissolve) with aggregate()
  2. merge non touching polygons by calculating convex hull hull(type = "concave_length")
  3. and simplify the geometry to remove some unecessary complicated geometries with simplifyGeom()

The perimeter is highly sensible to all theses steps (the area much less).

Figure 6: Exploration of how to merge sub-fields
Figure 7: Exploration of how to merge sub-fields

Number of sites with fields found in RPG sequence de culture


FALSE  TRUE 
    4   223

Summary of the area of fields (in \(ha\))

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 0.4956  2.3365  3.7666  4.6456  5.2915 29.0215

Summary of the perimeter of fields (in \(m\))

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
  307.4   641.1   849.3   916.3  1067.6  2566.4
Figure 8: Interactive map with the sites and the corresponding field from RPG sequence de culture.

TO BE DISCUSSED:

  • are the areas and perimeters within expected range? what about the shape of the fields?
  • how do we deal with issues identified visually, e.g. BIOHME_43_2020 (multiple fields + road)?

BD Haie and buffer

Let’s explore the dataset from BD Haie v2 mars 2024 from https://geoservices.ign.fr/bdhaie (1.5Gb compressed, 6.8Gb). BD Haie v2 is derived from images of the period 2020-2022 which is a better fit to our dataset than v1 based on images from 2011-2024.

Visualization

Let’s zoom in an area around Pleine-Fougères to visualize the hedges in BD HAIE (Figure 9).

Figure 9: Interactive map with the hedges and the sampled fields around Trans-la-Foret

Length of hedges

Summary of the length of hedges in the sampled field (m)

     Min.   1st Qu.    Median      Mean   3rd Qu.      Max.      NA's 
   0.9686   50.4023  129.8334  183.5141  246.4039 1068.4764        35

Summary of the density of hedges in the sampled field (km/ha)

    Min.  1st Qu.   Median     Mean  3rd Qu.     Max.     NA's 
0.000156 0.014081 0.036067 0.048862 0.068179 0.235265       35
Figure 10: Relation between field size and length of hedges within the field

Some fields (N=35) don’t have hedges around the field (e.g. Figure 11)

Figure 11: Example of a field with no hedges within the cropping area

Buffers

Let’s create four different buffers: 10m, 500m, 1000m, and 1500m (Figure 12) and calculate the density of hedge within the buffer (Figure 13).

Figure 12: Representation of the different buffer size
Figure 13: Relationships among hedge density calculated with different buffer sizes (1500m, 1000m, 500m, 10m, and no buffer).

The edge densities calculated with a large buffer (>500m) are highly correlated, especially the buffers 1000m and 1500m. The local hedge density might be better estimated with a buffer of 10m around the field (it will solve the issue seen in Figure 11).

TO BE DISCUSSED:

  • are the hedge density within expected range?
  • from Figure 9, it seems that not all hedges are mapped, but maybe only the interesting one. Are the results representing what we want? or should we complete with another dataset?